representation of that entry.

**NOTE**: It is more efficient to first `tf.bitcast` your tensors into
`int32` or `int64` and perform the bitcount on the result, than to feed in
8- or 16-bit inputs and then aggregate the resulting counts.
  }];

  let arguments = (ins
    TF_IntTensor:$x
  );

  let results = (outs
    TF_Uint8Tensor:$y
  );

    [FTRL paper](https://research.google.com/pubs/archive/41159.pdf).
*   `Optimizer.__init__` now accepts a `gradient_aggregator` to allow for
    customization of how gradients are aggregated across devices, as well as
    `gradients_transformers` to allow for custom gradient transformations (such
    as gradient clipping).
*   Improvements to Keras preprocessing layers: